Soft-Money Compute Strategy
The soft-money compute strategy describes how AI companies obtain computing power without paying for it with their own money: instead of cash, company shares, credits, or discounts flow. This allows young companies to use huge computing capacities, but ties them closely to their financial backer.
Anyone who wants to build a powerful AI system today needs a huge amount of computing power. This means thousands of specialized computer chips running for months on end. That quickly costs hundreds of millions of euros, and young companies don’t have that kind of money in the bank. The soft-money compute strategy is the usual answer to this: computing power isn’t paid for with cash, but with something else. The provider of the computers, for example, receives shares in the company, i.e. a portion of future profits. It’s called “soft money” because no hard cash changes hands.
Why compute time has become a second currency
In the AI industry, computing power is currently scarcer than money. Investors are pouring billions into new companies, but the right chips are sold out. Whoever has access to large data centers therefore has the upper hand. This is exactly what corporations like Microsoft, Amazon, Google, or Nvidia take advantage of.
For a start-up, the advantage is obvious. It can immediately train at large scale without first spending years raising capital. Without such arrangements, many well-known AI labs likely wouldn’t exist today. The strategy is thus what made the competition possible in the first place.
The price for this is independence. Anyone who sources their computing power from a single corporation cannot simply switch providers. Critics speak of a dependency that looks like a partnership on paper. Antitrust authorities in the US and the EU are now scrutinizing such arrangements more closely.
What is actually being exchanged in such deals
The most common form is the exchange of shares for compute time. A cloud provider makes computing capacity worth several billion available. In return, it receives a percentage of the company. Not a single euro is transferred, yet large sums appear on both sides' balance sheets.
A second variant is credits, as they’re technically called. The start-up receives a credit balance that it can only redeem with this one provider. A well-known pattern is the cycle: a chip manufacturer invests money in an AI company, and that company then uses it to buy chips from the very same manufacturer. Experts call this circular financing, because the money ends up right back where it came from.
A common misconception is that such deals cost nothing. They cost ownership. Anyone who gives up ten percent of their company permanently forgoes ten percent of all future profits. Compute time is used up once training is finished, but the equity stake remains forever.
How to spot such arrangements in the news
Reports about billion-dollar investments in AI companies very often involve such arrangements. When a headline mentions ten billion dollars, it’s worth reading the fine print. It often turns out that a large portion of it flows as cloud credits. The best-known case is the relationship between Microsoft and OpenAI; similar arrangements exist between Amazon and Anthropic.
This is also relevant for investors. When chip manufacturers co-finance their customers, their revenue looks strong even though part of the money originates from their own coffers. That’s why financial media discuss whether the AI industry’s revenue figures seem overly optimistic. The word “bubble” comes up regularly in this context.
In everyday life, you notice little of all this. You only feel it indirectly when a chatbot runs preferentially on a particular cloud. But for the question of which AI companies will survive in the long run, the compute question is often more important than the quality of the models.